AWS now offers over 30 AI and ML services. The breadth is impressive — but it also means teams frequently pick the wrong tier, over-engineer solutions with SageMaker when a pre-trained API would suffice, or under-invest when custom models are actually needed.
This guide maps every AWS AI service into a decision framework: three tiers based on how much control (and responsibility) you want, with clear guidance on when each service fits.
┌─────────────────────────────────────────────────────────────────┐
│ Tier 3: PRE-TRAINED AI APIs (No ML expertise needed) │
│ Rekognition | Textract | Comprehend | Polly | Transcribe | │
│ Translate | Lex | Personalize | Forecast | Fraud Detector | │
│ Kendra | HealthScribe | CodeWhisperer │
├─────────────────────────────────────────────────────────────────┤
│ Tier 2: GENERATIVE AI PLATFORM (Amazon Bedrock) │
│ Foundation Models | Agents | Knowledge Bases | Guardrails | │
│ Flows | Prompt Management | Fine-tuning | Model Evaluation │
├─────────────────────────────────────────────────────────────────┤
│ Tier 1: CUSTOM ML PLATFORM (Amazon SageMaker AI) │
│ Training | Inference | MLOps | Pipelines | Feature Store | │
│ Model Monitor | Ground Truth | Autopilot | Canvas │
├─────────────────────────────────────────────────────────────────┤
│ Infrastructure: EC2 (GPU/Trainium/Inferentia) | S3 | ECS/EKS │
└─────────────────────────────────────────────────────────────────┘
Rule of thumb: Start at Tier 3 (cheapest, fastest). Move to Tier 2 if you need generative AI or customization. Drop to Tier 1 only when you need full model control.
Use when: You need to train custom models on your own data, have ML engineers on staff, or require full control over the model lifecycle.
| Feature | What It Does |
|---|---|
| SageMaker Studio | |
| Unified IDE for ML development (notebooks, experiments, pipelines) | |
| Training | |
| Managed distributed training on GPU/Trainium instances with spot pricing (up to 90% savings) | |
| Inference | |
| Real-time endpoints, batch transform, serverless inference, and async inference | |
| Autopilot | |
| AutoML — automatically trains and tunes models from tabular data with zero code | |
| Canvas | |
| No-code ML for business analysts (point-and-click model building) | |
| Pipelines | |
| CI/CD for ML — automated training, evaluation, and deployment workflows | |
| Feature Store | |
| Centralized feature repository shared across training and inference | |
| Model Monitor | |
| Detects data drift, model quality degradation, and bias in production | |
| Ground Truth | |
| Data labeling service with human-in-the-loop workflows | |
| Clarify | |
| Model explainability and bias detection | |
| JumpStart | |
| Pre-trained model hub (deploy open-source models with one click) |
Use when: You want to build generative AI applications without training models from scratch. Bedrock provides serverless access to foundation models plus the building blocks for production AI applications.
| Provider | Models |
|---|---|
| Amazon | |
| Nova (Micro, Lite, Pro, Premier), Nova Canvas (images), Nova Reel (video), Nova Sonic (speech), Titan (Embeddings, Text, Image) | |
| Anthropic | |
| Claude 4 Sonnet, Claude 4 Opus, Claude 3.5 Haiku | |
| Meta | |
| Llama 4 Scout, Llama 4 Maverick | |
| Mistral AI | |
| Mistral Large, Mistral Small | |
| Cohere | |
| Command R+, Embed | |
| DeepSeek | |
| DeepSeek-R1 | |
| Stability AI | |
| Stable Diffusion XL, Stable Image | |
| AI21 Labs | |
| Jamba Instruct | |
| Luma | |
| Ray 2 (video generation) |
| Feature | What It Does |
|---|---|
| Agents | |
| Autonomous AI agents that plan, execute multi-step tasks, and call APIs. Support memory retention and multi-agent collaboration | |
| Knowledge Bases | |
| RAG (Retrieval-Augmented Generation) — connect your documents, databases, and data sources to ground model responses in facts | |
| Managed Knowledge Base | |
| Fully managed RAG (GA June 2026) — no vector DB provisioning, auto-scaling, multimodal ingestion | |
| Guardrails | |
| Content filtering, denied topics, PII detection, word filters, hallucination prevention — safety layer for any model | |
| Flows | |
| Visual workflow builder for chaining prompts, models, and data sources into applications | |
| Prompt Management | |
| Version, test, and optimize prompts across models | |
| Fine-tuning | |
| Customize foundation models with your data (continued pre-training or instruction tuning) | |
| Model Evaluation | |
| Compare models on quality, latency, and cost for your specific use case | |
| AgentCore | |
| Runtime infrastructure for deploying agents with tool use, orchestration, and observability |
Use when: Your problem maps to a well-defined AI task (vision, language, speech). No ML expertise required — just API calls.
| Service | Use Case | Pricing Model |
|---|---|---|
| Amazon Rekognition | ||
| Face detection, object recognition, content moderation, celebrity recognition, PPE detection | Per image/minute of video | |
| Amazon Textract | ||
| Extract text, tables, and forms from scanned documents (OCR on steroids) | Per page | |
| Amazon Lookout for Vision | ||
| Industrial defect detection from images (manufacturing quality control) | Per image |
| Service | Use Case | Pricing Model |
|---|---|---|
| Amazon Comprehend | ||
| Sentiment analysis, entity extraction, topic modeling, PII detection from text | Per unit of text | |
| Amazon Translate | ||
| Real-time language translation (75+ languages) | Per character | |
| Amazon Kendra | ||
| Intelligent enterprise search (understands natural language queries across documents) | Per index + queries |
| Service | Use Case | Pricing Model |
|---|---|---|
| Amazon Transcribe | ||
| Speech-to-text (real-time and batch, supports medical and call analytics variants) | Per second of audio | |
| Amazon Polly | ||
| Text-to-speech (lifelike voices, SSML support, neural voices) | Per character | |
| AWS HealthScribe | ||
| Medical conversation transcription with clinical note generation | Per minute |
| Service | Use Case | Pricing Model |
|---|---|---|
| Amazon Lex | ||
| Build chatbots and voice bots (powers Alexa). Intent recognition, slot filling, dialogue management | Per request | |
| Amazon Q | ||
| AI assistant for business (Q Business) and developers (Q Developer). Connects to enterprise data sources | Per user/month |
| Service | Use Case | Pricing Model |
|---|---|---|
| Amazon Personalize | ||
| Real-time recommendations (products, content, search ranking) | Per recommendation | |
| Amazon Forecast | ||
| Time-series forecasting (demand, inventory, capacity) | Per forecast | |
| Amazon Fraud Detector | ||
| Online fraud detection using ML models trained on your data | Per prediction | |
| Amazon DevOps Guru | ||
| ML-powered operational anomaly detection for AWS resources | Per resource/hour | |
| Amazon CodeGuru | ||
| Automated code reviews and application performance recommendations | Per lines analyzed |
| Service | Use Case | Pricing Model |
|---|---|---|
| Amazon Q Developer | ||
| AI coding assistant — code generation, debugging, transformation, security scanning | Per user or free tier | |
| Amazon CodeWhisperer | ||
| (Now part of Q Developer) Real-time code suggestions in IDE | Included in Q Developer |
| Service | Best For |
|---|---|
| EC2 P5/P4 instances | |
| GPU training with NVIDIA H100/A100 (highest performance) | |
| AWS Trainium (Trn1/Trn2) | |
| Cost-effective custom model training (up to 50% cheaper than GPU) | |
| AWS Inferentia (Inf2) | |
| High-throughput, low-cost inference at scale | |
| EC2 UltraClusters | |
| Massive distributed training (thousands of GPUs) | |
| Elastic Inference | |
| Attach fractional GPU to any EC2 for cost-efficient inference |
START
│
├── Do you need generative AI (text/image/video generation)?
│ ├── YES → Amazon Bedrock
│ │ ├── Need RAG? → Bedrock Knowledge Bases
│ │ ├── Need agents that call APIs? → Bedrock Agents
│ │ ├── Need safety controls? → Bedrock Guardrails
│ │ └── Need custom model? → Bedrock Fine-tuning or SageMaker
│ │
│ └── NO → Continue ↓
│
├── Does a pre-trained API solve your problem?
│ ├── Extract text from documents? → Textract
│ ├── Analyze sentiment/entities? → Comprehend
│ ├── Detect objects in images? → Rekognition
│ ├── Convert speech to text? → Transcribe
│ ├── Convert text to speech? → Polly
│ ├── Translate languages? → Translate
│ ├── Build a chatbot? → Lex (rules-based) or Bedrock (AI-powered)
│ ├── Search enterprise docs? → Kendra or Q Business
│ ├── Product recommendations? → Personalize
│ ├── Time-series forecast? → Forecast
│ └── Detect fraud? → Fraud Detector
│
└── Need to train a custom model on your data?
└── YES → Amazon SageMaker AI
├── Have ML engineers? → SageMaker Studio + Pipelines
├── No-code needed? → SageMaker Canvas or Autopilot
└── Deploy open-source model? → SageMaker JumpStart
| Criteria | Bedrock | SageMaker |
|---|---|---|
| Primary use | ||
| Generative AI applications | Custom ML model lifecycle | |
| Model ownership | ||
| Use provider models (no training) | Train/own your models | |
| Infrastructure | ||
| Fully serverless | Managed instances (you choose size) | |
| Customization | ||
| Fine-tuning, RAG, prompt engineering | Full training from scratch | |
| Scaling | ||
| Automatic (pay per token) | Manual endpoint scaling | |
| ML expertise needed | ||
| Low (application developer level) | High (ML engineer level) | |
| Best for | ||
| Chatbots, copilots, content generation | Recommenders, forecasting, custom NLP/CV |
The overlap: SageMaker JumpStart can deploy foundation models too. But Bedrock adds the application layer (agents, RAG, guardrails) that SageMaker doesn't provide natively.
The AWS AI stack is deep, but the decision framework is simple:
The most common mistake is jumping to SageMaker when Bedrock or a pre-trained API would solve the problem in days instead of months. Start simple, measure, and escalate only when the simpler tier doesn't meet your requirements.
Alpesh Kumbhare is an AWS Architect at Atos, specializing in AWS infrastructure automation and cloud AI solutions. Connect on LinkedIn.
source & further reading
dev.to — original article
Using HTTP QUERY in Laravel Without Making Your API Harder to Maintain
I Built an AI That Analyzes Chinese A-Share Stocks Using a Three-Pillar System — Here's What It Caught Today
Building a unified OpenAI-compatible gateway for 16 Chinese LLMs: architecture notes on routing, payments and ops